AI is so basic
Ignore the AI tech-hype cycle and focus on the longer-term, more transformative software trends.
Basic has two definitions - the formal: "simple and not complicated, so able to provide the base or starting point from which something can develop". Case in point: starting your sauce with onions and garlic. Basic, but not pretending to be anything else - you need to layer your additional flavours on top to make something special.
The second definition became prominent fairly recently, "boring and not unusual or surprising in any way". Basic in this context is often paired with "b*tch", ironically celebrating a love of, say, Taylor Swift.
Most algorithms powering our entertainment feeds (Instagram, YouTube, Spotify) and generating display ads stall in basic bitch mode. Only TikTok's audience graph has developed this base into something that tends to provide more interesting and varied recommendations.
My YouTube feed is full of different versions of what are essentially the same videos; my Spotify auto-generated playlists full of songs I've already listened to. Retargeting invariably involves products you've already bought following you incessantly around the internet. There are very few bold recommendations coming from the code. The blunt message is: never leave! Stay with us and watch more, shop more, consume MORE!
The brutal war for attention and digital advertising revenue plays a significant role in feed recommendations' endless, basic sameness. More lateral suggestions involve risk - branching out from the basic may mean your user switching to another platform, incrementally eating into your revenue. The safe play is to remove the risk, stay with the basic, and hope that your userbase doesn't get bored (and to be fair, there's often room in my life for another Radiohead listen or another video about fuzz pedals).
The continued basic unimaginative bluntness of recommendation algorithms explains the extreme hype around ChatGPT and other generative AI software. Using generative AI feels vaguely magical, a step closer to the sci-fi experience people inevitably expect when they hear the phrase AI. The outputs of tools like ChatGPT and Midjourney appear more nuanced and complex than anything we've seen before (consciously leaving aside the algorithms we don't see, such as those that make Google Translate almost as good as a translation agency).
And the tech-hype narrative driven by Silicon Valley and further fuelled by the media is that these algorithms will only get better as they get fed more data and receive more user feedback. People see a self-fulfilling loop that provides a potential pathway to artificial general intelligence (AGI). In other words, ChatGPT could end up directly spawning the AIs from your favourite sci-fi - Hal 9000, Holly, that cute kid from the Kubrick/Spielberg movie.
And there's no doubt that if we zoom out to look at the long-term trends, the artificial intelligence we use daily, whether seen or unseen, has improved exponentially over the past 15 years. But much of this improvement has been slow and unnoticed.
That jars with the VC-driven, media-powered tech hype cycle, which needs to see regular updates, launches and new features shipped. Every year, we want a new iPhone; we need new features and formats to keep the cottage industry of "Is x a y killer?" and "Ten ingredients for your x strategy" articles alive on LinkedIn. In economic terms, as
explains in his excellent piece "Probable events poison reality" we need to see the short-term return on investment to deem a technology viable.Based on this rationale, the tech industry must see continual growth from AI tools - whether existing software or new launches. Waiting five years for the next big leap won't keep the technological hype cycle spinning. We need new versions of ChatGPT and Midjourney; better generative AI for video creation.
But progress isn’t always linear and short-term. It’s taken years for generative AI tools to become genuinely useful. And in addition to more time, generative AI services need more training material to continue innovating and shipping new versions. More words, posts, images, interactions; more of everything. Neither of these challenges (requiring more time and requiring more data) are easily solvable.
If we look at the desire for more data, there's the increasingly urgent and thorny subject of IP and rights for the databases LLMs (large language models) access. Sarah Silverman is suing both OpenAI and Meta over copyright infringement; Reddit monetised its API and sent its community into meltdown to get value from its archive. Any business that hasn't already issued specific rules to employees about providing proprietary information to services like ChatGPT has potentially already given away IP for training data. People are wise to the worth of their words and work. There’s likely to be less data available for training, not more.
Secondly, as this fantastic joint piece of reporting by New York Magazine and The Verge highlights, a large and largely unseen human element goes into training LLMs like ChatGPT. Companies like OpenAI outsource data training to companies like Amazon's Mechanical Turk, allowing them to tap into a virtual army of workers who bid to complete small tasks. It's a fascinating piece I encourage you to read; the TL;DR version is that this work is complex and time-consuming, but the pay can be as low as $1 an hour for some annotators in Kenya. And the kicker is that some of these workers have turned to ChatGPT to help maximise the amount of money they can make.
Thirdly and finally, we're already starting to see news stories and reports about the rise of "next-generation content farms" building websites filled with AI-generated junk text and imagery. Currently, the numbers are small - according to MIT Technology Review, 25 new sites a week, on a base of around 200 or so (that it could find). But an estimated $13 billion of digital advertising revenue goes to existing "made for advertising" spam sites, 15% of spend. The unscrupulous can make decent money from using generative AI tools to spin out a few spam sites - then, you can easily see how online public spaces fill with auto-generated junk. More junk places a premium on the genuine interactions and conversations that have made tools like ChatGPT so powerful.
All three challenges mean less data - and the increased likelihood of AI tools being trained on AI-generated data - a bland, lowest-common-denominator cycle leading us back to basic.
All this isn’t to say that AI goes the way of the metaverse into the “dead trends” bucket. We must reframe how we think about AI - that catch-all term covering complex algorithms, machine learning, neural networks, and large language models.
The technologies that make up modern AI aren’t new; they’ve been slowly developing and improving over time. That makes AI a proper bona fide trend - it fits the Bill Gates idiom that “we always overestimate the change that will occur in the next two years and underestimate the change that will occur in the next ten”. A tool like ChatGPT might hit a brick wall, and the tech-hype cycle will inevitably move on to something else. But the broader concept of humans outsourcing tasks to software and improving our outputs isn’t going anywhere.
Eventually, all the most important technologies reach a point where to quote John Lanchester in the LRB, they’re so high-functioning and ubiquitous, they’re not considered “tech”. The humble toaster is technology; we see it as a utility. Broadband. Smartphones. HD TVs. All technological marvels we take for granted, all happily retired from the tech-hype cycle, basic in the sense that they provide a platform for us to build on.
When the tech-hype crowd moves on from generative AI and the media gets bored of doomsday AI scenarios, don’t be lulled into inaction around assistive software and advanced algorithms. These tools will continue to play a big role in our lives, whether seen or unseen, and as comms professionals, we need to continually consider their impact on our clients and our work. We need to understand the basics and give ourselves the right platform to build from.